Mistral Releases Timestral Forecasting Models

Mistral has released Timestral, its first family of time-series foundation models. In a 15 October 2026 post on mistral.ai, the company said its largest model, at 209 million parameters, “ranks first on all three of the field’s main public benchmarks — GIFT-Eval, TIME, and fev-bench — outperforming the previous state of the art with only two-thirds of the model parameters.”

Those first-place rankings are Mistral’s claim. The company says results have been submitted to the GIFT-Eval and TIME leaderboards; a technical report with full results “will follow.”

What a time-series foundation model is

A time series is a sequence of values over time — sensor readings, electricity demand, stock volumes, website traffic. Forecasts built on those series drive capacity planning, energy dispatch, supply chains, risk and maintenance.

Classical tools such as ARIMA or linear regression are usually fitted to one series at a time and learn only from that series’ history. Many deep models (for example PatchTST-style architectures) can learn nonlinear patterns across related series, but they are often trained from scratch on each dataset, so knowledge does not transfer to a new plant or market.

A time-series foundation model (TSFM) borrows the large-model playbook: pre-train on a broad corpus of series, then forecast zero-shot on data the model has never seen — useful when labelled history is scarce. Timestral is Mistral’s entry in that category, aimed at industrial and enterprise forecasting rather than chat.

The Timestral family

Mistral ships four sizes sharing one architecture and training recipe: 5M, 28M, 123M and 209M parameters. Features listed:

Probabilistic forecasts with 21 quantiles
Multivariate targets
Past-only and known-future covariates (for example production schedules or weather forecasts)
Context length of 32,768 time steps
Licence: Modified MIT
Deployment on-prem, VPC or air-gapped

Mistral says Timestral does not need a GPU at inference. On an 8-thread server CPU, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model — company latency examples.

On size comparisons, Mistral claims its 28M model beats Chronos-2 (about four times larger) on GIFT-Eval and TIME, and its 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters. Treat those as Mistral’s evaluations.

How Mistral says it was trained

The post sketches several techniques to be detailed in the forthcoming technical report: Decoupled Horizon Forecasting (so forecast tokens read full history without treating targets as missing values); covariate-aware attention that keeps targets causal while known-future covariates are visible; mostly synthetic pre-training followed by post-training on mixed real and synthetic data with a longer context; and a pinball loss over 21 quantiles approximating CRPS.

Adaptation runs through Mistral Forge — zero-shot use, fine-tuning on proprietary data, architectural extensions and air-gapped pipelines. A GitHub repository is published for running Timestral on local data.

Industrial example Mistral gives

Mistral walks through a plant with thousands of sensors: multivariate forecasts that jointly model vibration, temperature and pressure; quantile bands for risk-aware alerts; categorical covariates such as machine state; and known-future covariates such as planned shutdowns. It also sketches pairing Timestral with language-model agents for root-cause drafts when an anomaly fires — a product narrative, not a named customer case study.

Why it matters

Foundation models for language dominate headlines; forecasting still runs much of the physical economy. A small, CPU-friendly TSFM from a major open-weight lab widens Mistral’s product surface beyond LLMs — after moves such as acquiring New Material and releasing Large 4 Preview. Until leaderboards publicly list the submitted scores and the technical report lands, the “ranks first on all three” line should stay attributed to Mistral.

Caveats

Benchmark leadership is Mistral’s evaluation pending leaderboard listing. No technical report yet. Synthetic-heavy training means production teams will still want domain fine-tuning. Latency figures are company examples on a stated CPU setup.